AI & Automation

Buy or Build an AI Agent: A Cost Method You Can Fill In Yourself

November 11, 2026 · 6 min read · Etzal Group Team

In one meeting, the head of IT argues for building a customer service agent in-house because "we will own it and it will fit us exactly". The head of operations argues for subscribing to a product because "it works today". Neither brings a number. The decision to buy or build an AI agent can be made with a worksheet that takes an afternoon, and the exercise usually changes at least one person's mind.

Define the job first

Both routes are priced against the same job. Write it in five lines.

  • What the agent does. For example: answer order status and return questions on chat, and create a return request in the ERP.
  • Volume. Conversations or tasks per month now, and expected in two years.
  • Systems it must touch. ERP, CRM, payment, messaging.
  • Quality bar. What share must be right, and what must never happen.
  • Where data may go. Any restriction on location or third parties.

If these five lines are still being argued, neither buying nor building is ready.

The worksheet: six lines, three years

Fill in each line for each route. Use your own figures. The descriptions say what belongs in the line.

Cost lineBuyBuild
1. SetupConfiguration, connectors, content preparation, vendor onboarding feeDesign, development, integration, testing
2. Recurring platform feeSubscription per seat, per agent or per resolved conversationHosting, tooling licences, monitoring tools
3. UsagePer-conversation or per-action fees beyond the planModel usage fees at your volume
4. PeopleAn administrator who owns content and settingsAn owner, plus engineering time for maintenance
5. ChangePaying the vendor for custom requests, or waiting for the roadmapDeveloping new capabilities, migrating model versions
6. ExitExporting content and history, re-implementing elsewhereUsually low, you own it

Sum each column for year one, year two and year three. Three years matters, because build costs are front-loaded and product costs are level or rising with volume.

How to estimate the hard lines

Usage, when buying. Read the price page carefully. Many products now charge per resolved conversation or per action. Multiply by your realistic volume in year two, not the trial month. Ask what counts as a resolution and who decides.

Usage, when building. Estimate the amount of text in and out per conversation, including the instructions and retrieved documents sent with every request, multiply by the provider's current price and by volume. Then add a margin, because real conversations run longer than test ones. Measure on a prototype if you can. Guessing here is where build budgets go wrong.

People. Neither route runs itself. A product needs someone to keep content current and review conversations. A build needs that person and engineering time. As a planning assumption, set aside a meaningful fraction of the build cost each year for maintenance, then refine it with your team's own experience.

Change. List three things you expect to want in year two. For buying, ask the vendor whether each is possible and at what cost. For building, estimate the days.

Four questions money does not answer

1. Is this a differentiator? If the agent does what every company in your sector needs, a product will be good and getting better. If it embodies something particular to how you operate, a product will keep pulling you toward the average.

2. How deep is the integration? Products connect well to common systems and shallowly to unusual ones. If the value depends on acting inside your custom ERP, find out exactly what the product's connector can do, by testing it.

3. What do your data rules allow? If data must stay in a particular country or environment, the list of eligible products may be short.

4. Can you operate it? A built agent needs an owner who understands evaluation, monitoring and model changes. If nobody will hold that role, building creates a system that slowly decays.

Typical outcomes

The worksheet tends to point in recognisable directions.

  • Low to moderate volume, common use case, standard systems: buy. The build cost is never recovered.
  • High volume, per-resolution pricing: the product's usage line grows with success. Building, or negotiating a different price, starts to make sense.
  • Deep integration with custom systems, or a process that is your advantage: build, or build the integration layer and use a product for the conversation.
  • Strict data location rules: build on infrastructure you control, or choose among the few products that comply.
  • No one to own it: buy, and assign an administrator.

The middle route

The choice is rarely all or nothing.

  • Product plus custom actions. Use a product for the conversation, and build the tools it calls in your systems. You own the valuable part, the integration, and rent the commodity part.
  • Framework build. Build on an agent framework and a model provider's interface, not from first principles. Most custom work today is this: assembly, integration, evaluation and guardrails.
  • Buy now, build later. Start with a product to learn real volumes and real questions. Keep your content, test questions and conversation data portable. Revisit in a year with evidence.

Avoiding lock-in on either route

Whatever you choose, keep these in your own hands:

  • Your knowledge content, in a format you can move.
  • Your test set: the questions and expected answers that define good.
  • Your conversation history, exportable.
  • Your integrations, behind interfaces that do not depend on one vendor.
  • A written description of the agent's job and limits.

With those, switching routes later is a project. Without them, it is starting over.

Mistakes on both sides

Building:

  • Budget ends at go-live.
  • No evaluation set, so nobody can tell whether a change improved things.
  • Underestimating usage cost.
  • Building what a product already does well.

Buying:

  • Pricing by the demo month, not year two volume.
  • Assuming the connector does what the brochure says.
  • Not checking data terms.
  • No administrator, so content goes stale.

A one-week way to decide

  1. Day 1. Write the job in five lines.
  2. Days 2 and 3. Trial two products on twenty real conversations each. Note what they cannot do.
  3. Day 3. Ask a builder for a rough estimate against the same five lines.
  4. Day 4. Fill in the worksheet for three years.
  5. Day 5. Answer the four non-cost questions and decide: buy, build, or the middle route.

We build agents and we also tell clients when a product will serve them better, as part of AI Consulting. If you would like a second opinion on a worksheet you have filled in, contact us.

Summary

Define the agent's job in five lines. Fill in six cost lines for buying and building over three years, with honest estimates for usage, people and change. Then ask whether the agent is a differentiator, how deep the integration goes, what your data rules allow and who will operate it. Consider the middle route, and on any route keep your content, test set, history and integrations portable.

Frequently asked questions

Is it cheaper to buy an AI agent product than to build one?

Usually at the start and at low volume, yes. A product has no build cost and a predictable fee. A custom build costs more upfront and can cost less per use at high volume, or deliver something no product offers. The crossover depends on your volume, your integration needs and how much of the product's price is per-use.

What costs do people forget when building an AI agent?

The costs after launch: model usage fees, monitoring, evaluation, updating knowledge, migrating when the model provider retires a version, and the person who owns the system. These recur every year. A build budget that ends at go-live understates the true cost considerably.

Can we start with a product and build later?

Yes, and it is often the sensible order. A product lets you learn what the agent must actually do, what volume is real and where the product falls short. Keep ownership of your own content, test questions and conversation data so that a later build starts from evidence instead of from guesses.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.